By Matt Smith
European businesses are racing to meet EU AI Act obligations, but most are overlooking the place where their greatest governance exposure actually lives.
European boards and compliance teams are mapping AI systems, classifying risk levels, and building the documentation trails the EU AI Act demands. It is serious, necessary work. But most of it is focused on the AI tools organisations formally procure and deploy.
As Matt Smith, Field CTO, EMEA at Island, explains, the far harder problem of governing how employees actually use AI, every day, in real time remains largely unaddressed.
A widening governance gap
AI adoption inside most organisations has outpaced the frameworks meant to govern it. Employees are using generative tools, coding assistants, and increasingly, autonomous agentic AI in their daily workflows. Most of these tools fall outside of formally provisioned systems. While desktop-based apps like Claude Desktop are becoming more prevalent, they most often operate through a browser tab. Sensitive business data moves into these tools routinely, often without any awareness of where it is retained or who can access it.
Regulation is beginning to catch up. The EU AI Act represents the most significant attempt yet to place formal obligations around AI transparency, auditability, and data handling on European organisations. Some of its headline deadlines have already slipped, reflecting the difficulty of governing AI in practice. But the direction of travel is clear, and the compliance questions it raises – who is using which AI tools, what data are they handling, and how to demonstrate oversight – are ones every board should be asking regardless of where any specific deadline lands.
The reality is that most organisations cannot answer them. Compliance efforts tend to focus on AI systems that have been formally approved and procured. What they rarely account for is the shadow AI problem: the steady accumulation of unsanctioned tools that employees adopt independently, accessed through a browser session that leaves no audit trail and sits outside the reach of most security and governance architectures.
Agentic AI tools – those that take actions autonomously inside business systems on a user’s behalf rather than simply responding to prompts – make this more urgent still. They move data at a speed and scale that makes after-the-fact detection largely redundant.
Work moved, but the security architecture didn’t follow
To understand why the governance gap exists, it helps to look at how the working day has actually changed. For most employees, the browser is no longer a tool they use to access work. It is where work happens. SaaS platforms, cloud collaboration tools, financial systems, HR applications, and now AI tools all primarily live inside browser tabs.
Security architecture has not made the same journey. The controls most organisations depend on were built for a world of corporate networks, managed devices, and applications sitting behind a firewall. That model has not reflected reality for years, but it still underpins how most businesses approach security. The result is a significant blind spot: most controls can establish whether a user is permitted to reach an application, but they have no view of what happens once the user is inside it.
That gap matters for reasons that go well beyond compliance. When an employee copies a customer record into a personal email, pastes financial data into an external AI platform, or downloads a sensitive report to an unmanaged device, none of those actions necessarily triggers an alert. The access was authorised. The behaviour was not governed.
Organisations have responded by adding tools – data loss prevention solutions, cloud access security brokers, virtual desktop infrastructure, and VPN layers – that each are a reasonable response to a specific gap. However, together they form a stack that is expensive to run, difficult to manage, and still leaves the browser session itself largely ungoverned. The problem was never a shortage of tools. It was building on the wrong foundation.
The browser as the governance layer
The logical response to a governance gap that lives inside the browser is to move the enforcement layer there too. Not as another tool added to an already crowded stack, but as a fundamental shift in where security and oversight operate.
An enterprise platform that governs all browser activity embeds governance directly into the workspace, including an enterprise browser built specifically for organisational use and consumer browsers through a browser extension.
The distinction matters. Rather than attempting to intercept data after it has moved or block access before it is granted, real-time controls apply at the presentation layer, between the screen and the end user. A file downloaded to a personal device, a paste action directed at an unsanctioned AI tool, and a screenshot of confidential data become governable in real time, without disrupting the legitimate work happening around them.
This also changes what zero trust – the security principle of verifying identity and intent continuously rather than assuming it – can deliver in practice. Most implementations check identity once, at the point of login, and apply relatively fixed controls from there. An endpoint-native model assesses session context throughout the working day, adjusting policy as circumstances change. The user’s role, the data in motion, the tools being accessed, and the device in use all inform what is permitted at any given moment, rather than what was permitted at nine in the morning.
For the governance obligations that AI adoption is creating, this visibility is transformative. Security and compliance teams can see which AI tools are in use across the organisation, what data is being shared with them, and enforce policy accordingly – building the audit trail that regulators are beginning to require, and that boards should already be demanding.
Hybrid working and personal device use fit naturally into this model too. Organisational data is fully governed, while personal browsing remains genuinely private. The long-standing tension between employee privacy and organisational oversight resolves without either side having to concede ground.
Conclusion
The question European boards should be asking is not simply which AI tools their organisation has approved. It is whether they have the required visibility into how AI is being used across the working day, by whom, and with what data. For most organisations, the honest answer is “no.”
Regulation is creating pressure to change that, but the stronger argument is a practical one. Organisations carrying the cost and complexity of a sprawling security stack, while still lacking governance at the point where risk actually materialises, are neither well protected nor well positioned. The foundation needs to move to where work has already moved.
For most organisations, that journey starts with the browser.
